In February 2024, the King County Prosecuting Attorney's Office in Washington State issued a directive that landed differently in every police department it touched. The directive barred prosecutors from filing charges in cases where the arresting officer's report had been drafted by an AI tool without specific disclosure of that fact and without a documented verification process. It was not a ban on AI tools. It was a ban on undisclosed, unverified AI drafts entering the criminal justice pipeline as sworn accounts. The directive made the King County Prosecuting Attorney one of the most visible figures in a debate that the public-safety community had not yet fully engaged: the debate about what obligations come with AI-assisted documentation, and whether the legal system, as currently designed, can fully trust what comes out of an AI drafting tool. That debate is not resolved. And an officer who cannot articulate the concerns on both sides of it is not ready to use these tools responsibly.

The King County Objection: What It Actually Says

The King County directive is important to understand precisely, not just by its headline. The concern the prosecutor was raising was not that AI tools are inherently unreliable or that technology has no place in police work. The concern was more specific and more legally grounded: a police report is evidence. It is disclosed to the defense, examined by the prosecutor, and potentially read in open court. The chain of custody for that document, including the question of who created it and how it was verified, matters in the same way that chain of custody for physical evidence matters.

When an AI drafts a report narrative and an officer adopts it without a documented review process, several evidentiary questions become unanswerable. Did the officer verify every factual claim against the body-worn camera (BWC, the recording device officers wear to document encounters) footage? Were there details in the draft that the officer accepted without cross-checking against the actual record? If the AI-generated narrative contains a date, a description, a sequence of events, or an attributed statement that differs from what the footage actually shows, at what point was that discrepancy identified, and by whom?

The prosecutor's concern is not theoretical. In the adversarial setting of a criminal proceeding, the defense attorney's job is to probe exactly these questions. Under Brady v. Maryland (1963, requiring prosecutors to disclose evidence favorable to the defense) and Giglio v. United States (1972, requiring disclosure of evidence that could impeach the credibility of a government witness, including a police officer), the prosecution has a constitutional obligation to disclose material that affects the reliability of evidence in the case. An AI-drafted report whose generation process is undisclosed and unverified creates a potential Brady-Giglio problem because the defense cannot evaluate whether the document is the officer's independent recollection and observation or a model's statistical output that the officer passively accepted.

The King County position, read carefully, is not "do not use AI." It is "if you use AI, you must disclose it, you must document the review, and the report must be the officer's verified sworn account, not an unreviewed AI product." That is a more demanding standard than many departments were meeting when the directive was issued, and it is a standard this program treats as the appropriate floor, not an extreme position to debate.

What the EFF Is Warning About

The Electronic Frontier Foundation (EFF), the civil-liberties organization that has spent decades analyzing the intersection of technology and civil rights, has published detailed analysis of AI tools in law enforcement. Its concerns cluster around several distinct issues, each worth understanding on its own terms rather than as a unified opposition position.

Transparency and the Black-Box Problem

The EFF's first category of concern is transparency. Many AI tools used in public safety, including the models that generate report narratives from body-camera audio, are proprietary systems. Their training data, their weights, their specific decision architecture, and their failure modes are not publicly disclosed. When a report is generated by a tool whose internal workings are opaque, the defense attorney who wants to challenge the accuracy of the report cannot examine the tool that produced it. The model cannot be cross-examined. Its error rate in the specific context of the incident cannot be independently verified from public data.

This is not a frivolous concern. In the evidence world, when an expert relies on a proprietary method to reach a conclusion, there are established procedures for challenging the reliability of that method. The Daubert standard (the federal legal test, derived from Daubert v. Merrell Dow Pharmaceuticals, 1993, that governs whether scientific expert testimony is reliable and admissible) and its state equivalents give courts a framework for evaluating whether a scientific or technical method is sound enough to be presented to a jury. AI report-drafting tools are producing content that is being placed into sworn legal documents, but the "methodology" of those tools has never been subjected to Daubert-style scrutiny as a class. The EFF is flagging this gap before the adverse cases that will eventually force courts to address it.

Accuracy Disparities Across Populations

The EFF's second major concern is accuracy disparity. AI transcription and summarization tools are trained on datasets that may not be equally representative of all speakers, accents, dialects, and speech patterns. Research on speech recognition accuracy has documented that some commercial transcription systems perform less accurately for speakers of certain racial and ethnic backgrounds, for non-native English speakers, and for speakers with regional accents or speech patterns that are underrepresented in training data. If a report drafting tool that works from BWC audio transcription performs less accurately for encounters involving certain communities, then the reports it generates for those encounters will contain more errors than reports generated from encounters with other populations.

The implications are significant. If an AI-assisted report contains more gap-fill details or more inaccurate characterizations in encounters involving speakers the model was less accurately trained on, those errors are not randomly distributed across the population. They are systematically distributed in ways that correlate with race, ethnicity, and language. In a system already under scrutiny for bias, a tool that compounds transcription errors for specific communities is a civil-rights concern, not a technical footnote.

No public safety agency that is deploying AI transcription-based tools should do so without specifically testing and monitoring the accuracy of those tools across the full demographic range of the population the agency serves. The EFF's warning is that this testing is not routinely happening, and the disparity problem is therefore going unmeasured and unaddressed.

Vendor Lock-In and the Sole-Source Contract

The EFF and other technology-accountability advocates have also raised concerns about the procurement structure that is emerging around AI in public safety. Major vendors are offering bundled, multi-year, sole-vendor contracts that combine body cameras, cloud storage, drone integration, and AI tools into a single package. These contracts, which have reached values of approximately $45 million and terms of up to ten years in documented procurements, create a vendor relationship that is difficult to exit. When an agency's body cameras, evidence management platform, cloud storage, and AI report-drafting tool are all from the same vendor, the cost of switching vendors for any one component becomes prohibitively high. The vendor relationship becomes effectively permanent.

The concern here is about accountability and leverage. When a vendor's AI tool produces a problematic output, an agency that is locked into a multi-year sole-source contract has limited leverage to demand changes. When a vendor updates the model in ways that change its behavior, the agency may not even be informed, because model updates are often deployed silently as software-as-a-service improvements. The agency's criminal justice information services (CJIS, the Criminal Justice Information Services Security Policy that governs the handling, transmission, and storage of criminal justice data, with requirements that stay with the agency regardless of who the vendor is) compliance obligations stay with the agency even when the underlying technology is managed by the vendor. A vendor failure, a breach, a model change that introduces new error patterns, or a business decision to discontinue a product does not relieve the agency of its legal obligations under any reports that were generated by that product.

The Constitutional Framing: Brady, Giglio, and What AI Changes

The most important reason to understand the legitimate concerns about AI in police report writing is not vendor accountability or civil-liberties advocacy. It is the constitutional framework that governs every criminal case. Brady and Giglio are not procedural technicalities. They are the constitutional floor below which criminal proceedings cannot fall without becoming fundamentally unfair.

Brady v. Maryland established that the prosecution must disclose evidence that is favorable to the defense and material to guilt or punishment. A detail in a police report that is inaccurate, fabricated by an AI tool, and used to support a prosecution is not just an error. If the accurate version of that detail would have been favorable to the defense, it is a Brady violation. The officer who adopted the AI draft without verification, the prosecutor who filed charges based on the report, and the agency that deployed the tool without a disclosure policy are all implicated.

Giglio v. United States established that the prosecution must disclose evidence that could be used to impeach the credibility of a government witness, which in most cases includes the arresting and reporting officer. An officer whose report was drafted by AI without disclosure has a potential Giglio problem in every case that report touches: the defense can argue that the officer's credibility as a reporter of facts is compromised by the undisclosed AI-mediated process, and that the jury should know how the report was generated. If that disclosure comes out during cross-examination rather than in pre-trial discovery, the damage to the officer's testimony, the case, and the department's reputation is maximized.

The constitutional framing reorders the risk calculus. The question is not whether AI drafting tools are useful (they are), or whether they save time (they do), or whether the 82 percent time reduction figure is credible (it is a serious benchmark). The question is whether the agency's deployment of those tools meets the constitutional standards that govern what happens to the output in court. An agency that deploys AI drafting without disclosure policy, without documented verification, and without a CJIS-compliant data-handling framework is not just taking a technology risk. It is taking a constitutional risk on every case in which an AI-assisted report is used.

The Automation Bias Problem

Beyond the legal and constitutional framing, there is a behavioral risk that is worth naming directly: automation bias. Automation bias is the well-documented human tendency to over-rely on automated outputs and to reduce the critical scrutiny applied to information that a machine has produced. Research across aviation, medicine, finance, and other high-stakes domains consistently shows that when a trusted system provides a confident output, human reviewers are less likely to notice errors in that output than they would be if they had produced the analysis themselves.

In the context of AI police report drafting, automation bias poses a specific risk. An officer who has used a drafting tool repeatedly and found it generally accurate will tend to scan the draft rather than verify it against the footage. The draft that "looks right" at a glance gets signed. The detail that the model invented, which looks exactly like a detail the officer would have written, passes the scan because the officer is looking for gross errors rather than performing a systematic verification against the record. The result is that the officer's review becomes a rubber stamp rather than a verification, and the tool's failure modes enter the sworn report unchecked.

Departments that have been using AI drafting tools for more than a few months have anecdotally reported a pattern where early-adoption officers perform rigorous verification and later-adoption officers become less thorough over time as familiarity with the tool increases. This is automation bias at work. It is not a moral failing; it is a documented behavioral pattern. The professional response to it is not to distrust the tool, but to build verification processes that are systematic and independent of the officer's subjective sense of whether the draft "looks right." The footage-grounded verification pass described in later lessons in this program is designed precisely to counteract automation bias by giving the officer a specific, methodical checklist rather than a general instruction to "review the draft."

Reading the Concerns as a Professional

An officer who has read this far might reasonably ask: if the concerns are this substantial, should AI drafting tools be used at all? The answer this program gives is the same answer it would give about any powerful tool: the tool should be used when the professional using it understands its risks, applies the appropriate verification discipline, and operates within a disclosure framework that makes its use defensible in court. That is a higher bar than passive adoption. It is also a lower bar than prohibition.

The King County prosecutor did not ban AI. The EFF did not call for a moratorium on body cameras. The constitutional framework established by Brady and Giglio does not make AI tools inadmissible. What the concerns establish is the price of admission: disclosure by design, verification as a professional discipline, and accountability that stays with the human author regardless of what tool generated the first draft.

An officer who can articulate why King County raised its objection, what the EFF is specifically warning about, and how Brady and Giglio apply to AI-assisted documentation is an officer who can use these tools responsibly, explain their use to a prosecutor or a defense attorney, and answer the deposition question without flinching. That officer is also the one who can tell a chief or a city council member the same true thing: "We understand the risks, we have the verification standard, and we have the disclosure policy." That competence is what this program is building.

An officer who cannot articulate the risk cannot responsibly use the tool.

Key Takeaways

• The King County (WA) Prosecuting Attorney barred AI-written police reports that lacked documented disclosure and verification. The directive was not a ban on AI; it was a demand that AI-assisted reports meet the evidentiary standards required for criminal prosecution.
• The EFF has raised three specific categories of concern about AI in police report writing: transparency and the black-box problem (proprietary tools cannot be cross-examined), accuracy disparities across populations (transcription tools may perform less accurately for certain communities), and vendor lock-in through bundled multi-year sole-source contracts reaching $45 million and up to 10 years.
• Brady v. Maryland and Giglio v. United States are the constitutional frame for AI in public safety. An AI-generated inaccuracy that would have been favorable to the defense is a Brady violation. An undisclosed AI drafting process that undermines an officer's credibility as a fact reporter is a Giglio problem.
• CJIS (Criminal Justice Information Services) Security Policy compliance obligations stay with the agency regardless of which vendor manages the underlying technology. A vendor breach or model update does not transfer the agency's legal obligations.
• Automation bias, the human tendency to reduce critical scrutiny of trusted automated outputs, is a documented behavioral risk in AI report review. Building systematic verification processes counteracts automation bias more reliably than relying on individual officer vigilance.
• The concerns raised by King County, the EFF, and the constitutional framework do not prohibit AI use; they establish the price of admission: disclosure, documented verification, and human accountability for the sworn product.
• An officer who can articulate the concerns is more professionally capable with AI tools than one who cannot. That articulation is the difference between responsible deployment and liability exposure for the officer, the agency, and every case the tool touches.
• Accuracy disparities in AI transcription tools across racial, ethnic, and linguistic communities are a civil-rights concern. Agencies deploying transcription-based drafting tools should test and monitor accuracy across the full demographic range of the population they serve.